AI-Assisted Biparametric MRI Surveillance of Prostate Cancer: Feasibility Study

PURPOSE: To evaluate the feasibility of automatic longitudinal analysis of consecutive biparametric MRI (bpMRI) scans to detect clinically significant (cs) prostate cancer (PCa). METHODS AND MATERIALS: This retrospective study included a multi-center dataset of 1539 patients who underwent bpMRI (T2 + DWI) between 2014--2020, of whom 105 patients underwent at least two consecutive bpMRI before biopsy without pathologically confirmed csPCa prior to follow-up. A deep learning prostate cancer detection model was developed and trained to produce a heatmap of all PIRADS>=2 lesions across baseline and current studies. The aligned heatmaps for each patient's baseline and current examination were used to extract differential volumetric and likelihood features reflecting explainable changes between examinations. A logistic classifier was trained to predict from these features csPCa (ISUP>1) at the time of the current examination according to biopsy. A model trained on the current study only was developed for comparison. An extended model was developed incorporating clinical parameters (PSA density and age). Cross-validation was performed to assess the detection performance of the models on unseen data. The diagnostic performance of the best model was compared to the radiologist scores. Diagnostic accuracies are compared using likelihood ratio tests and ROC analysis. RESULTS: The model including baseline and current study (AUC 0.73 CI: 0.49 0.89) performed better than the current only model (AUC 0.70 CI: 0.42 0.86), and significantly (P=0.002) improved fit. Adding clinical variables further improved diagnostic performance (AUC 0.79 CI: 0.60 0.94). The extended surveillance model's performance was comparable to that of the radiologist (AUC 0.69 CI: 0.52 0.86). CONCLUSIONS: Our proposed AI-assisted surveillance of prostate MRI can pick up explainable, diagnostically relevant changes with promising diagnostic accuracy. CLINICAL RELEVANCE: Research on artificial intelligence that integrates longitudinal information from follow-up prostate MRI is lacking.